Active recall for Knowunity · iOS
Proof, not praise: voice recall for Knowunity
Knowunity's recall feature ended at a score. I designed what happens next.
- Usability testing
- Coded prototype
- AI-native process
- Role
- UX designer
- Timeline
- 2.5 weeks, September 2026
- Tools
- Figma, Claude Code, Next.js, Storybook
- Status
- Course project. Not shipped.
Context
Students revise, then can’t retrieve it in the exam, and nothing in the app tells them which is which
Problem 1
‘Almost there’ cannot tell whether students were misheard
Problem 2
The Summary screen names perfection and time to complete without giving a next step
Problem
Knowunity’s beta shows a verdict and a hint but never what the app heard, so a student who gets “Almost there” can’t tell whether they were misheard or genuinely partial, and a fair result can feel unfair. Its summary then names perfection and time to complete without giving a next step, so the student leaves with a judgment they can neither verify nor act on.
‘Almost there’ cannot tell whether students were misheard
The Summary screen names perfection and time to complete without giving a next step
Overview
Knowunity is a study app with 30M+ downloads. Its practice tools test recognition, like multiple choice and flashcards. None of them ask a student to pull an idea from memory and say it.
Users
Secondary school students and college students before exams. Flattering feedback sends them in over-confident, and punishing hints stops them asking for help.
Solution
An AI-powered voice experience that turns passive studying into active recall. Students explain concepts aloud, receive progressive hints when needed, and get a personalized breakdown of their strengths and knowledge gaps, along with a targeted study plan to improve weaker areas and build exam confidence over time.
Product film: a student allows the mic, answers aloud, climbs two hints, reviews their take, sees 'You explained 2 of 4 without help', then retries the topic they missed.
Role and approach
I owned the whole feature as the solo designer, from audit to coded prototype
What I owned: everything in a 2.5-week sprint with Knowunity, as the solo designer.
01
Beta audit
02
Figma design
03
Coded prototype
04
Usability testing
05
Rebuild after the test
Research
The beta graded students on the basis of answer but students lost track of the hint ladder
01
The beta’s grade misled.
It flattered a student who guessed and punished one who used help.
02
The hint ladder was invisible.
Participants lost track of where they were, and the reveal didn’t read as the last step.
03
The next action was unclear.
After the answer appeared, participants didn’t know whether to retry or submit.
I audited Knowunity’s beta before designing anything. Its intro promised scores “up to 20% higher on exams”. It ended with a grade and listed missed terms under a red “Needs practice”.
Beta audit. Drove three decisions: the take became a choice, the wait stopped looking like “sent”, and praise became a count.
Competitive analysis shows how four apps handle recall, misses, repetition and the summary. Drove the decision to show the student’s answer beside the model answer, as Anki and Quizlet do.
Exploration
I explored three interaction directions to increase Active Recall adoption, targeting 25% activation among students and 75% task completion among those who started.
Voice recall: the loop, its states, and where it breaks
- Screen
- State
- Edge case
- The payoff
Getting there
Screen
Choosing feature from AI Chat
The student opened the app. This is the moment they would chose the feature.
Screen
Glance over material
Users has the option to look through the reading materials before they activate explain out loud.
State
Mic permission
The primer sits before the OS dialog. Designed the denied path, not only the granted one.
The recall loop
Screen
Prompt
"Explain X in your own words." Progress is visible. So is skip.
State
Idle
Mic ready. Three ways out: speak, type, skip.
State
Recording
Stopping is sending. Cancel discards the take without costing the turn.
State
Processing
Transcribing, then judging. Under four seconds. The hardest state in the whole feature.
State
Result
Pass, partial, or miss. Judge generously: a false "wrong" costs more here than anywhere else in the app.
When it doesn't go clean
Edge
Partial answer
Hint that nudges without giving it away. Re-attempt. Second hint. Re-attempt. Then reveal.
Edge
Can't speak right now
Train, library, shared room. One tap to type. Same judging, same hints, no penalty.
Edge
Nothing usable heard
Silence, or a transcript that clearly is not an answer. Re-record without losing the turn.
Edge
Leaves mid-session
Progress saves, returning resumes. Not a drop-off until genuinely abandoned.
Ending it
Screen
Summary
How many they explained unaided. Hinted, revealed and skipped stay distinct from each other.
Payoff
A 2 columns with correct answer and answer given by you screen
The felt signal of "I actually know this now." If overconfidence costs nothing, this screen is flattery.
Screen
What happens next
Continue, or re-test what they missed. Recall works through repetition over time, not one clean session.
Usability test
A think-aloud usability test surfaced eight problems
“I do not know from the reveal question that it was distinct from the other hint.”
Entry point, Recall loop and summary screen.
All eight things the test surfaced
01
The reveal screen lacked hierarchy. Participants didn’t know it was the final step.
02
The next action was unclear: retry, continue or submit, especially after the reveal.
03
Hints needed a visible progression, so a student knows how much help they’ve had and what comes next.
04
Reviewing and retrying were mixed together. The mic should stay off while the answer shows, until the student chooses to retry.
05
Participants wanted to see how much was right: a progress circle or percentage, on the verdict and the final result.
06
The summary should say what to do next, per topic, not only list concepts to review.
07
Participants wanted to compare with other students’ answers, to find a better way to explain it.
08
Half-remembered concepts should come back, asked a new way, like Duolingo, with a note on why.
Key decision
I rebuilt the hint ladder so every rung looks like the step it is
Option A
Three identical hint dots
The reveal still offers the mic.
Option B (chosen)
Named rungs: Hint 1, Hint 2, Hint 3, Reveal
- The current rung lights up
- Hint 3 gets a lightbulb and a gold box
- Hints appear inside the verdict card
Clip from the product film: a student gives a partial answer to "What was appeasement, and why did it fail?", scores 65% with Hint 1 inside the verdict card, answers again, and climbs to 75% with Hint 2 as the next rung lights up.
Final designs
Each screen answers one question, and every turn has a way out
1 / 6The reveal hands the mic back. No mic while you read; Try now opens a fresh recorder.
Screen recording of the prototype: from the home screen, the student opens World War II, chooses Explain out loud, answers, sees "What you said" with Retry and Continue, reveals the answer, and waits while it evaluates.
AI-native process
I used AI to build and audit the prototype, and kept the design decisions myself
What stayed mine
Every product decision came from me, the tests or the brief. I accepted or rejected each critic finding myself.
I decide
Brief, decisions log, rules
Claude Code builds
Screens from my Figma and the component library
Critic agents grade
11 scorecards, hard gates
The system I set up
Project rules
Every value from the design tokens, only components from the Storybook library, sentence case, Knowie never speaks, and every screen keeps a way out.
Coded prototype
I built a Next.js prototype of 22 recall states from my Figma screens, with Claude Code in VS Code.
Component library
32 components in Storybook. Every screen uses them.
Critic agents
Four critic agents (craft, UX, design system, ambition) and a spec reviewer.
Rubric and hard gates
Six dimensions and four hard gates, such as 48 px tap targets and 4.5:1 contrast.
Product film
Claude Code rendered the film from the real components. It can't show anything the prototype doesn't do.
Semantic colour tokens in Storybook
Each with a note on the job it does.
The HintLadder component
Its Figma source, the reason for each state, and where it departs from Figma.
Outcome
What I can claim: a tested, iterated prototype, not a shipped result
- Recall states in the coded prototype
- 22
- Components in the Storybook library
- 32
- Critique rounds by critic agents
- 11
What exists today. Nothing shipped, so I don’t claim a metric. A pilot would measure Knowunity’s targets: 25% of students who see the feature start it, and 70% of those finish it. These are targets, not results.
Reflection